Usage limits
The daily limits on each plan, what counts toward them, and what happens when you reach one.
Each plan has four daily limits. They reset at 00:00 UTC every day, and there are no overage charges: when a limit is reached, that kind of work stops until the reset.
The limits on each plan
| Per UTC day | Free | Pro | Enterprise |
|---|---|---|---|
| Inference requests | 1,000 | Unlimited, subject to fair use (250,000) | Sized to your usage |
| AI fallbacks | Not included | 500 | Sized to your usage |
| Training minutes | 15 | 60 | Sized to your usage |
| Generated examples | 2,000 | 20,000 | Sized to your usage |
These numbers come from the same settings that enforce the limits, so they are always current. Every plan allows any number of projects. See Pricing for what else each plan includes.
What counts
| Limit | What uses it |
|---|---|
| Inference requests | Every request your endpoints answer, from your app, the intent bar, Try it or Automated tests. |
| AI fallbacks | Each request your model couldn't answer that was sent to the AI fallback. Requests with the AI fallback switched off don't use it. |
| Training minutes | The time your training runs take. A run starts only while you have minutes left, and its minutes are added when it ends. |
| Generated examples | Training pairs written with Gemini on the Generate step, including the examples of requests your tools can't answer. In the Terse style only the templates Gemini writes count, so Terse produces many more pairs per example counted. |
The Profile page shows these as Inference requests, AI fallbacks, Training minutes and Gemini examples, under Today's usage, with how much of each is left.
On Pro, inference requests are unlimited, subject to fair use. The number in the table is the fair-use ceiling, and the Profile page shows a count rather than a bar.
When you reach a limit
| Limit | What happens |
|---|---|
| Inference requests | Your endpoints answer 429 with Daily request limit reached for this project. and quota_exceeded: true, until the reset. |
| AI fallbacks | Requests your model can't answer come back with no calls, needs_fallback: true and quota_exceeded: true. Requests it can answer are unaffected. |
| Training minutes | Starting a run fails with Daily training-minutes limit reached. A run already in progress finishes. |
| Generated examples | The Generate step fails with Daily example-generation limit reached. Pairs already written are kept. |
Usage emails
Fernfly emails you when you pass 80% of a daily limit, and again when you reach it. One email covers every limit you passed. On Pro there is no 80% email for inference requests, since they aren't metered.
Grace credits
If you reach a limit and need to keep working today, ask for grace credits. They add to that limit for the rest of the UTC day.
- On the Profile page, under Today's usage, choose Request grace credits on the limit you reached. You can ask once a day.
- Or reply to the email telling you that you reached the limit.
We review each request and grant credits by hand, so they aren't instant.
Other limits
| Limit | Value |
|---|---|
| Training runs in progress | 3 across your account, 1 per project |
| Tools per generation | 100 |
| Tools for an intent bar | 30 |
| Keyless requests | 20 a minute per caller per project |
| Training data import | 10 MB and 20,000 pairs per file |
Revised